How to Filter Books with AI Without Wasting Your Time

You can finish a hundred books and still make little progress. The bottleneck is often selection, not reading speed: some books repeat what you already know, some stretch one idea across hundreds of pages, and others confidently package weak evidence as fact. Filtering should therefore happen before buying and reading, not afterward.
AI is a filter, not a replacement reader
A model cannot reliably judge a book from its title. Give it evidence: the publisher's description, table of contents, sample chapters, author background, and substantive reviews. The thinner the source material, the lower the confidence should be. A polished verdict without citations is still only a plausible guess.
The most important input is your profile. State your current level, goals, books you have already read, and topics you do not need repeated. An introduction may be valuable to a beginner and nearly useless to an experienced developer. There is no universal book score.
Criteria that produce useful reports
Ask for the same structured report for every candidate:
- what problem the book solves and who it targets;
- which chapters contain ideas that are new to you;
- whether major claims are supported by evidence;
- how much space goes to repetition and motivational stories;
- whether selected chapters are enough;
- which stronger alternatives cover the same goal.
A “filler percentage” is not a scientific measurement. Treat it as a comparison label within one consistent rubric. The evidence behind the score—chapters, claims, and contradictions—is more valuable than the number itself.
A practical pipeline for one hundred books
Build a table with title, author, reading goal, and available sources. Process books in small batches, save each report as Markdown, and sort the results into three queues: read fully, read selected chapters, or postpone. Flag every report where the model lacked enough evidence.
Then manually inspect the highest-ranked books. Read the contents page, a sample chapter, and several critical reviews. This takes minutes and catches persuasive hallucinations. AI can reduce a funnel of one hundred candidates to ten, but the final decision remains yours.
The goal is not an impressive counter of “processed” books. It is a small number of relevant books that you study carefully and turn into action.
How my own mistakes produced this system
I read roughly sixty books a year and used to treat weak books as the unavoidable cost of finding strong ones. Once my notes were in Obsidian, the difference became measurable. One book produced dozens of connected notes and decisions I applied to projects. Another produced three obvious quotations across four hundred pages. Ten mistakes like that consume eighty to one hundred hours.
Pseudoscience is worse than ordinary filler. A confident author, several success stories, and fluent prose can create the feeling of progress while the argument rests on “vibrations,” attraction, or health advice with no evidence. A bad book can cost more than time because you later have to unlearn its model of the world.
My first workflow was manual. I gathered tables of contents, reviews, and critical commentary in NotebookLM, described what I already knew, and evaluated one book at a time. It worked, but not for a backlog of one hundred titles. I moved the stable criteria into one prompt and wrote an asynchronous script that processes small groups concurrently and saves every report as Markdown.
What goes into my reader profile
“Recommend a good book” is almost useless. My profile contains my stack, experience, current projects, previous reading, and explicit exclusions. I do not need another Python syntax overview or a shallow Git introduction. Architecture, security, optimization, and practices I can test in a real codebase receive higher priority.
There is also a red-flag section: conspiracy claims, esotericism, unsupported medical promises, and endless stories without a method. A beginner needs a different profile. For someone entering Python, fundamentals are not filler. A copied prompt therefore produces a ranking for the wrong person; the criteria must describe your actual gap.
What the report contains
I want more than a score from one to five. The report identifies the audience, novelty for me, structure, repetition, evidence behind major claims, and useful chapters. Sometimes a complete book is not worth reading while two chapters solve a current problem. Only those chapters enter the queue.
The model also suggests alternatives, and this is where verification matters most. Popular does not mean better, and a model can invent a plausible title. I open the publisher page, verify the author, and read a sample. The result is a two-stage funnel: software cheaply removes obvious mismatches; human attention is reserved for the shortlist.
Markdown is deliberate. Reports remain readable without a proprietary service, live next to my Obsidian notes, work with ordinary search, and enter the same backup system. If an API disappears, the analysis remains mine.
Where the method fails
A niche technical book with no reviews, a new research monograph, and literary fiction do not fit this scoring model. A contents page cannot reveal the quality of prose. A language model should not perform medical or financial due diligence. Nor should every rejected book be counted as “saved time”: an unfamiliar position may be more valuable than material optimized for your existing preferences.
I therefore store a rejection reason and confidence level. When evidence is thin, a title moves to manual review rather than the trash. That small distinction is what turns an impressive AI demo into a dependable tool.
A minimal template you can reproduce
My script is not required for the first experiment. Choose five books and create columns for goal, prior knowledge, available evidence, useful chapters, red flags, confidence, and decision. Require the model to fill a field only from supplied material and state when evidence is absent.
Read one sample chapter from the highest- and lowest-ranked books. If the report cannot explain the actual difference, change the criteria rather than the model. Add ten more books only after the small set behaves consistently; automate the API last. This prevents a common failure: scaling a wrong prompt and producing one hundred equally confident, useless reports.
| Decision | Use it when |
|---|---|
| Read fully | High novelty, strong evidence, and direct relevance to current work |
| Read chapters | One section is valuable while the rest repeats known material |
| Postpone | Evidence is thin or the topic is not currently useful |
| Skip | Critical claims are unsupported or the level is a clear mismatch |